D e z e m b r o 2 0 1 2

410

CHOOSING AN OPTIMAL INVESTMENT STRATEGY: THE ROLE

OF ROBUST PAIR-COPULAS BASED

PORTFOLIOS

Beatriz Vaz de Melo Mendes Daniel S. Marques

Relatórios COPPEAD é uma publicação do Instituto COPPEAD de Administração da Universidade Federal do Rio de Janeiro (UFRJ)

Editor

Leticia Casotti Editoração Lucilia Silva Ficha Catalográfica Marisa Rodrigues Revert

Mendes, Beatriz Vaz de Melo. Choosing an optimal investment strategy: The role of robust

pair-copulas based portfolios / Beatriz Vaz de Melo Mendes, Daniel S. Marques. – Rio de Janeiro: UFRJ /COPPEAD, 2012.

23 p.; 27cm. – (Relatórios COPPEAD; 410)

ISBN 978-85-7508-097-9

ISSN 1518-3335

1. Finanças. 2. Cópulas (Estatística matemática). I. Marques, Daniel S. II. Título. IV. Série.

CDD – 332

Pedidos para Biblioteca: Caixa Postal 68514 – Ilha do Fundão 21941-972 – Rio de Janeiro – RJ Telefone: 21-2598-9837 Telefax: 21-2598-9835 e-mail: biblioteca@ coppead.ufrj.br Disponível em www.coppead.ufrj.br

Choosing an optimal investment strategy:

The role of robust pair-copulas based portfolios

Beatriz Vaz de Melo Mendesab1, Daniel S. Marquesa

a COPPEAD, Federal University at Rio de Janeiro, phone: 55 21 25989898, P.O. Box 68514, Rio

de Janeiro, RJ, 21941-972, Brazil.

b IM, Federal University at Rio de Janeiro. Av. Athos da Silveira Ramos 149, CT, P.O. Box

68530 21941, Rio de Janeiro, RJ, 21941-909, Brazil.

Abstract

This paper is concerned with the efficient allocation of a set of financial assets and its successful

management. Efficient diversification of investments is achieved by inputing robust pair-copulas

based estimates of the expected return and covariances in the mean-variance analysis of Markowitz.

Although the whole point of diversifying a portfolio is to avoid rebalancing, very often one needs to

rebalance to restore the portfolio to its original balance or target. But when and why to rebalance

is a critical issue, and this paper investigates several managers’ strategies to keep the allocations

optimal. Findings for an emerging market target return and minimum risk investments are highly

significant and convincing. Although the best strategy depends on the investor risk profile, it

is empirically shown that the proposed robust portfolios always outperform the classical versions

based on the sample estimates, yielding higher gains in the long run and requiring a smaller number

of updates. We found that the pair-copulas based robust minimum risk portfolio monitored by a

manager which checks its composition twice a year provides the best long run investment.

Keywords: Pair-copulas; Optimal financial portfolios; Robust estimation; Rebalancing.

JEL Classifications: C01, C19,C32, C51, C58, G11.

1Corresponding author (Permanent address): Beatriz Vaz de Melo Mendes, phone: 55 21 22652914,

Rua Marquesa de Santos, 22, apt. 1204, Rio de Janeiro, 22221-080, RJ, Brazil. Email: beatriz@im.ufrj.br.

1

1 Introduction

Financial institutions and portfolio managers are primarily concerned with the efficient

allocation and monitoring of sets of financial assets. Periodic portfolio rebalancing, aiming

to restore the investment back to its desired target risk and return, is a crucial step in the

process of controlling risk. Commonly asked questions are how often a portfolio should

be rebalanced, and which would be the best indicators of changes in the global economy

or in the balance among the component assets.

Efficient diversification of investments based on the mean-variance analysis of Markowitz

(1952) is widely used by institutional investors. Statistically, the resulting efficient frontier

just relies on the estimates of the expected return and covariance matrix, and the sample

estimates are the usual inputs. However, the statistical good properties of the sample

estimates are attached to the highly improbable assumption of multivariate normality.

A better characterization of the data underlying multivariate distribution will provide

more reliable estimation of the efficient frontier. This means we must know not just the

marginal univariate series behavior and their correlations, but their whole d-dimensional

probability distribution. This may be accomplished by modeling the data through pair-

copulas (Frigessi and Bakken (2007); Min and Czado (2008); Berg and Aas (2008); Fischer

et al. (2008)).

A pair-copula construction is just a hierarchical decomposition of a multivariate copula

into a cascade of bivariate copulas. Since an appropriate copula function can be found

for any type of association — linear, nonlinear, ranging from perfect negative to perfect

positive dependence — one can expect the model to truthfully represent the data at hand.

Estimation takes place at the level of the two-dimensional data, therefore avoiding the

famous curse of dimensionality.

The analysis of financial data from emerging markets poses some specific challenges.

Atypical points in transaction prices (from non-confirmed unexpected news, market ma-

nipulation, and so on) distort classical statistical inference, corrupting the inputs to the

mean-variance algorithm. A distorted correlation matrix and inflated risk estimates will

provide misleading allocations. To handle deviations from the true underlying distribution,

robust methodologies are called for. We suggest to apply the robust estimates for pair-

copulas models, initially proposed in Mendes et al. (2007). For each parametric copula

family there exist a robust weighted minimum distance or a weighted maximum likelihood

estimator providing accurate estimates under contaminations. The robust portfolios are

obtained by inserting the robust pair-copulas based mean and covariance estimates in the

mean-variance Markowitz procedure.

2

Robust methods typically detect the pattern implied by the vast majority of the data,

providing more stable estimates. Robust allocations are resistant to unjustified sudden

fluctuations of the market, which are identified by the robust estimates as point contami-

nations. Therefore, robust portfolios are primarily designed for long run investments. We

note that the notion of “long run” may vary across markets and to account for changes in

the economy, some periodic rebalancing of the portfolio may still be needed. It is expected

from a robust investment to yield higher gains in the long run and to require a smaller

number of updates.

There is no universally accepted best strategy for portfolio management. Best strategy

will change with investor risk aversion, portfolio target return or standard deviation.

Among many others, we consider the popular strategy followed by institutional investors

that monitors a portfolios at an annual (or monthly) frequency and then rebalances only if

the allocation to an asset shifts more than some threshold (5%, 1%). We do not consider

additional factors when implementing the rebalancing strategies, such as trading costs or

cost of time spent which would reduce the return of the portfolio. However we keep track

of the rebalancing frequency of each manager and are able to draw some conclusions based

on their number of rebalancings.

Summarizing, in this paper we address both problems of composing and managing

portfolios, given a set of financial instruments. We do not address the issue of choosing

the component assets. We robustly estimate the data multivariate distribution using pair-

copulas obtaining the inputs which will define the robust efficient frontier. The trajectory

of a target return and the minimum risk portfolios will be managed by twelve managers

during a 2-years period of out-of-sample investigation. We use data from emerging mar-

kets because of the higher volatility of these stock markets and their greater potential

for interdependence with the major markets. More specifically, we use six-dimensional

contemporaneous daily log-returns from the most traded Brazilian stocks, due to Brazil’s

important position among emerging equity markets. The robust portfolios are compared

to their classical version based on the sample empirical estimates,

The contributions of this paper are three fold: (i) we introduce and investigate the per-

formance of pair-copulas based robust portfolios; (ii) we investigate 12 managing strategies

aiming to keep (or restore) the portfolio target, to guarantee the same risk aversion level;

(iii) we illustrate the ideas using Brazilian data.

Findings in the paper are striking and convincing. We found that despite the invest-

ment type, the robust methodology always outperform the classical version. We are able

to determine the best rule for restoring the portfolio to its original balance and keep the

allocations optimal. We show that the best strategy depends on the investor risk profile,

3

and that pair-copulas based robust minimum risk portfolios monitored by a manager which

checks its composition twice a year provides the best long run investment. Additional ex-

ercises are not provided because the aim of the paper is to find best portfolio composition

along with best long run managing strategy for a given set of financial instruments.

The remaining of this paper is organized as follows. In Section 2 we briefly consider the

classical mean-variance methodology for obtaining efficient portfolios, and briefly review

the definitions of pair-copulas and robust estimates. In Section 3 we define 12 strategies

for portfolio monitoring. In Section 4 we analyze two 6-dimensional data sets and assess

the performance of classical and robust target and minimum risk portfolios. In Section 5

we discuss and summarize the results. Section 6 provides the references.

2 Statistical Methodologies

Derived from simple mathematical terms relating the expected return and risk of a port-

folio, Markowitz’s optimization procedure (Markowitz, 1952) for obtaining the efficient

frontier can be considered the most widely used result of modern economics. Based on

the idea that one should diversify to reduce risk, for a given expected return, the portfolio

theory minimizes risk. To this end the theory considers not only the means and variances

of a set of securities, but also their covariances. However, the true data generating process

is unknown and the inputs for the mean-variance algorithm must be estimated.

2.1 Classical allocations

Markowitz mean-variance optimization is a quadratic optimization problem, whose classi-

cal inputs are the sample mean and the sample covariance matrix, the maximum likelihood

estimates (MLE) under multivariate normality. We call the portfolios obtained using the

classical inputs as classical portfolios.

One of the most important causes of limitation of the method in practice is the lack of

optimality presented by classical estimates in the financial environment. This problem was

studied by Klein and Bawa (1976), Jobson and Korkie (1980, 1981), Canela and Collazo

(2007), among others. Following Markowitz work, a large part the literature was devoted

to obtain reliable alternatives for the sample mean and the sample covariance matrix.

2.2 Robust pair-copulas based allocations

In the last decade, copulas have been widely used in finance (see for example, Embrechts

et al. (1999); Demarta and McNeil (2005); and Gatzert et al. (2008)). The reason for

this popularity is the inadequacy of the multivariate normal distribution when modeling

4

financial data. Any multivariate distribution may be decomposed on its marginal uni-

variate distributions and a copula, which contains all information about the dependence

structures linking the margins (Nelsen (2007); Joe (1997)). Therefore, modeling the data

through copulas allows one to obtain estimates of any characteristic of the distribution,

including Markowitz inputs, the mean vector and the covariance matrix. We call these

estimates as copulas- (or later pair-copulas-) -based estimates.

Consider a continuous random vector X1, . . . , Xd with joint cumulative distribution

function (c.d.f.) H(x1, . . . , xd) and marginal distributions F1, . . . , Fd. Sklar’s theorem

(Sklar, 1959) ensures that there exists a d-copula C such that for all (x1, . . . , xd) ∈

[−∞,∞]d

H(x1, . . . , xd) = C(F1(x1), . . . , Fd(xd)). (1)

Conversely, if C is a d-copula and F1, . . . , Fd are c.d.f.s, the function H defined by (1) is a

d-dimensional distribution function with margins F1, . . . , Fd. Furthermore, if all marginal

c.d.f.s are continuous, C is unique. A d-dimensional copula C is a d-dimensional c.d.f. on

[0, 1]d with standard uniform marginal distributions.

When C is absolutely continuous, taking partial derivatives of (1) one obtains

h(x1, · · · , xd) = c(F1(x1), · · · , Fd(xd))

d∏

i=1

fi(xi) (2)

for some d-dimensional copula density c. This expression is the basis for the usual two

steps inference approach, where firstly the marginal estimates are obtained, and then the

copula parameters are estimated (Joe and Xu (1996), Joe (1997)).

The equation (2) allows for tailored marginal modeling considering all characteristics

of each Fi, including the mean, standard deviation, skewness, kurtosis and any type of

short and long memory serial dependence, plus a search for the best fit for the dependence

structure through a large number of copula families that may be considered. This results

in flexible multivariate distributions well fitted to the data.

A copula function is invariant under monotone increasing transformations of X, making

copula-based dependence measures interesting scale-free tools for studying dependence.

An important copula-based dependence concept is the coefficient of upper tail dependence,

λU , defined as

λU = limα→0+

λU(α) = limα→0+

Pr{X1 > F−1

1(1 − α)|X2 > F−1

2(1 − α)} ,

provided a limit λU ∈ [0, 1] exists. If λU ∈ (0, 1], then X1 and X2 are said to be asymp-

totically dependent in the upper tail. If λU = 0, they are asymptotically independent.

5

Similarly, the lower tail dependence coefficient, λL, is given by

λL = limα→0+

λL(α) = limα→0+

Pr{X1 < F−1

1(α)|X2 < F−1

2(α)} ,

provided a limit λL ∈ [0, 1] exists. It follows that

λU = limu↑1

C(u, u)

1 − u, where C(u1, u2) = Pr{U1 > u1, U2 > u2} and λL = lim

u↓0

C(u, u)

u.

The coefficient of tail dependence measures the amount of dependence in the upper

(lower) quadrant tail of a bivariate distribution. In finance, it is related to the strength of

association during extreme events. The copula derived from the multivariate normal dis-

tribution does not have tail dependence. Therefore, if this copula is assumed for modeling

log-returns, for many pairs of variables it will underestimate joint risks.

More flexibility may be gainned by considering pair-copulas models. Pair-copulas is

an hierarquichal decomposition of a d-copula into a cascade of of potentially different

bivariate copulas. It was originally proposed by Joe (1997), and later discussed in detail

by Bedford and Cooke (2001, 2002), Kurowicka and Cooke (2006) and Aas et al. (2007).

The composing bivariate copulas may vary freely, with respect to choice of the parametric

families and parameter values. Therefore, all types and strengths of dependence can be

covered.

For large d, the number of possible pair-copula constructions is very large. Bedford

and Cooke (2001) introduced a systematic way to obtain the decompositions, the so called

regular vines. These graphical models help understanding the conditional specifications

made for the joint distribution. Two special cases are the canonical vines (C-vines) and

the D-vines. Again, the success of estimation procedure starts with good marginal fits (see

Frahm, Junker, and Schmidt, 2004), which typically pose no difficulties. MLE estimates

may be computed at both steps.

However, occasional atypical points may occur in finance, and they may corrupt the

classical estimates of the dependence structure. Mendes and Accioly (2011) proposed to

robustly estimate pair-copula models using the Weighted Minimum Distance (WMDE)

and the Weighted Maximum Likelihood (WMLE) estimates. They are based on either a

redescending weight function or on a hard rejection rule. Robust estimation of a pair-

copula model is also performed at the level of bivariate data, and details about the robust

estimates found in Mendes et al. (2007).

The WMDE minimize some selected weighted distance from the empirical copula, a

goodness of fit statistics. The WMLE result from a two-step procedure. In the first

step, outlying data points are identified by a robust covariance estimator and receive

zero weights, and in the second step the copula MLE are computed for the reduced data.

6

There are many high breakdown point estimators of multivariate location and scatter that

could be used in this preliminary phase. We use the robust Stahel-Donoho (SD) estimator

based on projections (Stahel, 1981 and Donoho, 1982) which is implemented in the free

R software. This affine equivariant estimator possesses high breakdown point, a bounded

influence function, all good robustness properties which are expected to carry over the

weights of the portfolios.

3 Portfolio Managing (Rebalancing: when and why)

Rebalancing an efficient portfolio means to update the data and re-evaluate the efficient

frontier obtaining a new set of weights defining the assets’ new allocations according to

the preferred risk exposure. Although involving costs, the task is performed because one

wants to bring the assets’ allocations back to the investor’s level of risk tolerance and

expected return.

The frequency of rebalancing is usually pre-determined and follows some rule. Many

rebalancing strategies may be formulated. We define 12 rebalancing strategies through

a list of explicit rules to follow, aiming to represent the various degrees of managers’

risk tolerance, some expressing a very rational line of thinking, others representing some

subjective approach. They determine how frequently the portfolio is monitored and also

which risk measure is being controlled. We do not address two issues. One, if a portfolio

should be rebalanced to its target or to a new allocation suggested by the economy. Two,

we do not consider rebalancing costs (transaction costs, taxes, time, labor, and so on),

which are difficult to quantify, but we do keep track of the number of rebalancing events.

The rules for rebalancing are:

Manager 1: Does nothing. Keeps the portfolio’s allocations unchanged.

Manager 2: Rebalances the portfolio weekly (5 business days).

Manager 3: Rebalances the portfolio monthly (21 business days).

Manager 4: Rebalances the portfolio every three months (63 business days).

Manager 5: Rebalances the portfolio every semester (126 business days).

Manager 6: Follows the Drawdown. If the portfolio’s Drawdown duration reaches three

days, he/she rebalances.

Manager 7: Follows the Drawdown. If the portfolio’s Drawdown duration reaches six

days, he/she rebalances.

7

Manager 8: Follows the Drawdown. If the portfolio’s Drawdown value is equal to or less

than the 0.25 quantile of the portfolio’s Drawdown empirical distribution, he/she

rebalances.

Manager 9: Follows the Value-at-Risk (VaR). He/she inspects the portfolio monthly (each

21 days) and rebalances if there were at least three returns smaller that the current

VaR0.01 in the last 100 days, or if the current VaR0.01 was not violated during the

last 300 days.

Manager 10: Re-evaluates the allocations monthly (each 21 days) and rebalances if for

any asset its updated allocation differs from its current allocation by 5% or more.

Manager 11: Re-evaluates the allocations monthly (each 21 days) and rebalances if for

any asset its updated allocation differs from its current allocation by 10% or more.

Manager 12: Re-evaluates the allocations every semester (each 126 days) and rebalances

if for any asset its updated allocation differs from its current allocation by 10% or

more.

In the analysis of Section 4 we investigate the performance of the managers when

monitoring a minimum risk and a fixed target portfolio. The goal is to find the best

rebalancing strategy which would provide the higher accumulated gain along a testing

period, while maintaining the main characteristic of each investment, namely, to possess

minimum risk, and to attain some target return, respectively. We assume that the assets

chosen to form the portfolio will be available during the validation period, and that they

will not be exchanged.

At each portfolio update, managers 6, 7, 8, and 9 also update the corresponding risk

measure. The VaR0.01 is computed as the 0.01-quantile of the empirical distribution of

the portfolio. The Drawdown is computed as the sum of losses occurring in a sequence

of negative daily returns. In other words, the drawdown is defined as the percentual

accumulated loss due to a sequence of drops in the price of an investment (Grossman

and Zhou (1993); Chekhlov et al. (2003)). It is a flexible risk measure collected over

non-fixed time intervals and provides a different perception of the risk and price flow of

this investment. Thus, for managers 6, 7 and 8 there is no periodicity in his/her behavior.

Rule followed by Manager 9 was somehow suggested by the data analyzed in Section 4,

which presented a period of high turbulence at the end of the estimation period, yielding

more extreme risk measures. So the rule tried to detect if the economy has changed either

back to a less volatile period, or even increasing to a more volatile one. We note that the

type I error probability in both cases are approximately 6% and 5%.

8

Managers 10, 11, and 12 follow rules which are popular among private investors and

also managers in financial institutions.

4 Empirical Analyzes

As stated in the Introduction, our objective in this empirical investigation is two-fold. For

a given data set, firstly we want to compare the performance of the robust and classical

portfolios (as defined in Section 2). Secondly, we want to assess the performance of the

twelve managers, finding the best rule for each portfolio type.

The daily returns were computed from 10 years of transaction prices on the six most

traded Brazilian stocks “Vale do Rio Doce”, “Petrobras”, “Usiminas”, “Banco do Brasil”,

“Eletrobras” and “Tim” provided by BOVESPA. We split the data in two periods, one

for estimation (in-sample 8-years period from June/22/2001 through June/19/2009) and

the 2-years out-of-sample period from June/20/2009 through June/24/2011.

We select two portfolio types for the investigations: the Minimum Risk, and a Fixed

Target. Since the choice of the target value is subjective, we decided to select a portfolio lo-

cated approximately at one third along the efficient frontier, for which all component assets

typically had a positive weight contribution. Actually, this target portfolio could represent

a popular choice for an investment in emerging markets, due to its low risk when compared

to the Maximum Risk portfolio, and to its interesting daily target return of 0.1040% cor-

responding to an annualized log-return of almost 30%. We note that the in-sample average

daily return for the six stocks are respectively (0.1164, 0.0935, 0.1600, 0.1159, 0.0442,−0.0097).

The portfolios computed at the end of the estimation period will be called baseline

portfolios. To obtain the classical portfolios’ allocations we compute the mean-variance

inputs using the classical sample estimates. To obtain the allocations of the robust port-

folios we carry on the two-steps estimation method, as follows.

Initially the unconditional distributions of each series is estimated. We fit by maximum

likelihood a skew-t distribution (Hansen, 1994) to the six 8-years series of log-returns, and

from the estimated c.d.f.s we obtain the pseudo uniform(0, 1) data. Estimates are not

shown here, but available to the reader by request. The marginal fits were carefully

checked since a poor fit would result in the probability integral transformed values not

being standard uniform or i.i.d.. As a consequence, any copula model would be mis-

specified.

9

V & PBB7(1.48, 0.94)

(0.477, 0.404)

P & UBB7(1.32, 0.81)

(0.424, 0.312)

AAAU

��

��

AAAU

U & BBBB7(1.31, 0.74)

(0.393, 0.302)

��

��

AAAU

BB & EN (0.496)

(0, 0)

��

��

AAAU

E & Tt(0.48, 24)

(0.007, 0.007)

��

��

V&U|P

t(0.282, 7)

(0.067, 0.067)

AAAU

P&BB|U

t(0.223, 9)

(0.030, 0.030)

AAAU

��

��

U&E|BB

BB1(0.271, 1.132)

(0.104, 0.155)

AAAU

��

��

BB&T|E

t(0.287, 10)

(0.031, 0.031)

��

��

V&BB|U,P

N(0.088)

(0, 0)

AAAU

P&E|BB,U

F(1.012)

(0, 0)

AAAU

��

��

U&T|BB,E

N(0.210)

(0, 0)

��

��

V&E|BB,U,P

t(−0.0223, 21)

(0.0001, 0.0001)

P&T|BB,U,E

t(0.102, 11)

(0.009, 0.009)

AAAU

��

��

V&T|BB,U,E,P

F(0.4661)

(0, 0)

Exhibit 1: D-vine decomposition. Best copula fits and their parameters estimates. The third row

inside the boxes gives the tail dependence coefficients (λL, λU). Notation in figure: Copulas: N:

Normal; t: the t-copula; F: Frank. Data: V: Vale; P: Petrobras; U: Usiminas; BB: Banco do

Brasil; E: Eletrobras; T: Tim.

Five unconditional and ten conditional bivariate copulas compose the D-vine. The

parametric copula families considered included the Normal, t-student, BB7, BB1, Clayton,

Gumbel, Tawn, Frank, and the product copula. These copula families cover all possible

combinations of values for the lower and upper tail dependence coefficients. In addition,

asymmetric dependence may be modeled by the Tawn copula, a non-exchangeable de-

pendence structure. To find the best copula fit, we compared the value of the penalized

log-likelihood (AIC), examined the pp-plots based on the estimated and the empirical

10

copula, and computed a GOF test statistic (Genest and Remillard (2005) and Genest et

al. (2007), Berg (2008)).

The D-vine decomposition, along with the robust copula fits with parameter estimates

and tail dependence coefficients (λL and λU), are shown in Exhibit 1. Pairs in the first level

are those possessing higher correlations. The robust fits reflect the pattern of the majority

of days and are expected to provide better inputs for the mean-variance algorithm and

therefore more consistent long-run portfolios.

Likewise in Mendes et al. (2010), we compute the rank correlations provided by the

pair-copula decomposition. The pair-copula-based robust estimates along with the skew-

t location and scale estimates provide the inputs for the mean-variance algorithm. We

run the long-only MV optimization algorithm and construct the classical and the robust

pair-copula-based efficient frontiers.

For evaluating the performance of the classical and the robust pair-copula based port-

folios we now assume that the four portfolios are managed by the twelve managers along

the 2-years validation period. The rules are followed and portfolios are updated according

to them.

Firstly, we fix the portfolio type (statisticalmethodology and investment objective) and

find the best rebalancing strategy for this financial instrument. To this end we (a) count

the percentage of time its accumulated gain under some rebalancing approach is higher

or equal than the other competitors; (b) test if the mean difference between accumulated

gains from every pair of managers is statistically different from zero, using a one-sided

robust nonparametric t-test (Wilcoxon rank sum test) at the 1% level.

We found that for Classical Target Portfolio, blindly rebalancing each 5 days (Manager

2) leads to consistently significantly lower accumulated gains. Keeping the allocations

unchanged for two years (Manager 1) is the second worst thing one can do (wins over only

Manager 2). For the Classical Target Portfolio the best strategy comes from Manager

10, which inspects the allocations each 21 days (every month) and rebalances only if any

allocation has changed by 5%. The proportions of times Manager 10 portfolio accumulated

gains is higher than the others is typically around 80%, and highly statistically significant,

see Figure 1. Figure 1 shows the differences between the accumulated gain from Manager

10 and all other managers. For this data set the resulting portfolio coincides with the one

from Manager 3 (which blindly rebalance each month).

For the Robust Target Portfolio the worst thing one can do is to blindly frequently

rebalance the portfolio each 5 days (Manager 2). This is in line with its long run investment

characteristic implied by the robust methodology. However, since the economy changes

and there is a target to follow, some type of re-allocations are needed, and the best manager

11

0 100 200 300 400 500

-2-1

01

2

M 10 - M 1

0 100 200 300 400 500

-2-1

01

2

M 10 - M 2

0 100 200 300 400 500

-0.4

-0.2

0.0

0.2

0.4

M 10 - M 3

0 100 200 300 400 500

-10

1

M 10 - M 4

0 100 200 300 400 500

-2-1

01

2

M 10 - M 5

0 100 200 300 400 500

-10

1

M 10 - M 6

0 100 200 300 400 500

-2-1

01

2

M 10 - M 7

0 100 200 300 400 500

-2-1

01

2

M 10 - M 8

0 100 200 300 400 500

-10

1

M 10 - M 9

0 100 200 300 400 500

-1.0

-0.5

0.0

0.5

1.0

M 10 - M 11

0 100 200 300 400 500

-2-1

01

2 M 10 - M 12

Classical Target Portfolio -- The difference: Manager 10 - All Managers

Figure 1: For the Classical Target portfolio figures show the differences between the accu-

mulated gains under Manager 10 and all other managers.

for the Robust Target Portfolio turned out to be Manager 9, which strategy is based on the

VaR value. Even though he/she inspects monthly, only 14 investment updates were done

(in contrast, for the Classical Target portfolio, the winner Manager 10 has rebalanced 23

times). Figure 2 shows the differences between the accumulated gain from Manager 9 and

all other managers for this portfolio.

On the other hand, when it comes to Classical Minimum Risk Portfolios, keeping

the allocations unchanged for two years (Manager 1) apparently came out as the best

rebalancing strategy. It lead to significant consistently higher accumulated gains when

compared to the gains from the remaining strategies. However, this result should be

looked at with care since the baseline allocations might not result in an efficient minimum

risk anymore. Figure 3 shows the updated efficient frontier after the two-years validation

period, and the position of the baseline classical minimum risk portfolio in the risk ×

return space.

Thus rebalancing is needed to keep the investment characteristic during its life time.

Manager 11 (which monitors the portfolio monthly with a 10% threshold) provided the

best maintenance strategy, see Figure 4. The second best manager is number 9 which uses

the VaR measure. From Figure 4 it is clear the superiority of these two strategies.

12

0 100 200 300 400 500

-2-1

01

2

M 9 - M 1

0 100 200 300 400 500

-2-1

01

2

M 9 - M 2

0 100 200 300 400 500

-1.5

-0.5

0.5

1.5

M 9 - M 3

0 100 200 300 400 500

-10

1

M 9 - M 4

0 100 200 300 400 500

-2-1

01

2

M 9 - M 5

0 100 200 300 400 500

-10

1

M 9 - M 6

0 100 200 300 400 500

-10

1

M 9 - M 7

0 100 200 300 400 500

-2-1

01

2

M 9 - M 8

0 100 200 300 400 500

-1.5

-0.5

0.5

1.5

M 9 - M 10

0 100 200 300 400 500

-1.5

-0.5

0.5

1.5

M 9 - M 11

0 100 200 300 400 500

-2-1

01

2

M 9 - M 12

Robust Target Portfolio -- The difference: Manager 9 - All Managers

Figure 2: For the Robust Target portfolio figures show the differences between the accumu-

lated gains under Manager 9 and all other managers.

Best strategy for rebalancing the Robust Minimum Risk portfolios came from Manager

12, whose strategy is to re-evaluate the allocations each 6 months and rebalance whenever

allocations change by 10% (see Figure 5). This is also in line with the long run characteris-

tic of robust procedures. We note that Manager 12 just rebalanced once and outperformed

Manager 1 which kept the allocations unchanged for two years. Interestingly, the second

best manager is Manager 9, which controls the Value-at-Risk of the portfolio.

Secondly, we compare the performances of the same type of classical and robust in-

vestments driven by their best managers. The upper plot in Figure 6 shows the differences

between the accumulated gains from the Robust Target & Manager 9 portfolio and the

Classical Target & Manager 10 portfolio. We observe that even though the target is the

same the robustly estimated pair-copulas based portfolios outperform the classical ver-

sion. The classical version was updated 23 times whereas the robust one only 14 times.

The lower plot in Figure 6 shows the outstanding superiority of the Robust Minimum

Risk investment & Manager 12 (rebalanced once) over the classical version & Manager 11

(rebalanced 9 times).

Following a suggestion from a referee, we now investigate the performance of the ro-

bustly estimated pair-copulas method combined with the proposed managers for different

13

Risk (standard deviation)

Exp

ecte

d R

etur

n

0.020 0.022 0.024 0.026 0.028

0.00

080.

0009

0.00

100.

0011

Baseline Classical Minimum Risk Portfolio

Figure 3: Updated efficient frontier and baseline Classical Minimum Risk portfolio.

financial instruments in the same country. Two stocks (Eletrobras and Tim) were replaced

by the series IMA-C and IMA-S, respectively a long-term inflation-indexed Brazilian trea-

sury bonds index and a floating rate Brazilian Treasury bill index, both computed by

the Brazilian Association of Financial Institutions, ANBIMA (www.anbima.com.br). The

IMA-S consists of the price changes of Letras Financeiras do Tesouro (LFT), which are

zero-coupon shorter term securities whose interest rate is compounded daily using the

average treasury repo market rates computed by the Brazilian Central Bank (the SELIC

rate) and that is paid according to this accruing in one single payment at the maturity

date.

We again split the data in two parts, a in-sample 8-years period for data estimation

and the 2-years out-of-sample period for models validation. The Fixed Target portfolio

now has a daily target return of 0.17%. The new series present positive in-sample average

daily return, respectively, 0.0786 and 0.0617. Their main characteristic is their much

smaller sample standard deviations, respectively, 0.204 and 0.054, when compared to

those computed for the remaining four stocks, respectively, 2.542, 2.452, 3.185 and 3.033.

The same estimation steps were followed and optimal portfolios’ weights computed. The

difference now between the classical and the copula based efficient frontiers is not so

dramatic, but the robust-pc based curve is still located above and at the left of the

14

0 100 200 300 400 500

-0.5

0.0

0.5

M 11 - M 1

0 100 200 300 400 500

-0.6

-0.2

0.2

0.6

M 11 - M 2

0 100 200 300 400 500

-0.6

-0.2

0.2

0.6

M 11 - M 3

0 100 200 300 400 500

-0.6

-0.2

0.2

0.6

M 11 - M 4

0 100 200 300 400 500

-0.6

-0.2

0.2

0.6

M 11 - M 5

0 100 200 300 400 500

-0.6

-0.2

0.2

0.6

M 11 - M 6

0 100 200 300 400 500

-0.6

-0.2

0.2

0.6

M 11 - M 7

0 100 200 300 400 500

-0.6

-0.2

0.2

0.6

M 11 - M 8

0 100 200 300 400 500

-0.6

-0.2

0.2

0.6

M 11 - M 9

0 100 200 300 400 500

-0.6

-0.2

0.2

0.6

M 11 - M 10

0 100 200 300 400 500

-0.6

-0.2

0.2

0.6

M 11 - M 12

Classical Minimum Risk Portfolio -- The difference: Manager 11 - All Managers

Figure 4: For the Classical Minimum risk portfolio figures show the differences between

the accumulated gains under Manager 11 and all other managers.

classical version. In particular, the weights defining the Minimum Risk portfolios are very

close, both methods allocating around 95% on the IMA-S, due to the obvious reason.

The target portfolios had almost all weight allocated to three components, the Vale and

Usiminas stocks and the IMA-C. We repeated the same strategies for evaluating and

comparing the performance of classical and pc-based portfolios.

We found that the best rebalancing strategy for the Classical Target Portfolio came

from Manager 6 which updates the optimal weights whenever three consecutive losses are

observed. The proportions of times Manager 6 Classical Target Portfolio accumulated

gains is higher than all others is around 84%, highly statistically significant.

For the Robust Target Portfolio the best manager turned out to be Manager 7, which

strategy is to update whenever there is a sequence of 6 negative returns. Only one update

was made, whereas for the Classical Target portfolio, the winner Manager 6 has rebalanced

19 times. For the Robust Target Portfolio he second best performance was provided by

Manager 4 which inspected at the end of 63 business days, having carried out seven

investment updates.

On the other hand, the Classical Minimum Risk Portfolio driven by Manager 12 pro-

vided higher accumulated gains than all others during time percentages varying between

15

0 100 200 300 400 500

-1.5

-0.5

0.5

1.5

M 12 - M 1

0 100 200 300 400 500

-10

1

M 12 - M 2

0 100 200 300 400 500

-10

1

M 12 - M 3

0 100 200 300 400 500

-10

1

M 12 - M 4

0 100 200 300 400 500

-10

1

M 12 - M 5

0 100 200 300 400 500

-10

1

M 12 - M 6

0 100 200 300 400 500

-10

1

M 12 - M 7

0 100 200 300 400 500

-10

1

M 12 - M 8

0 100 200 300 400 500

-10

1

M 12 - M 9

0 100 200 300 400 500

-10

1

M 12 - M 10

0 100 200 300 400 500

-10

1

M 12 - M 11

Robust Minimum Risk Portfolio -- The difference: Manager 12 - All Managers

Figure 5: For the Robust Minimum Risk portfolio figures show the differences between the

accumulated gains under Manager 12 and all other managers.

67% and 97%. This manager has inspected the allocations every semester using a fluctu-

ation margin of 10%, but has performed only one update. For the minimum risk portfolio

we observed that the weights distribution along the validation period was very stable,

explaining the second place been occupied by three managers which performed zero up-

dates, namely managers 1, 7, and 8. We note that the minimum risk portfolio has allocated

approximately 95% weight to the low risk IMA-S.

Best strategy for rebalancing the Robust Minimum Risk Portfolio came from Manager

10 , which re-evaluates the allocations monthly and rebalances if the updated allocation

changes by 5% or more for any asset. The number of updates carried out was 9.

Finally, we compare the performances of the same type of classical and robust invest-

ments driven by their best managers. Likewise Figure 6, Figure 7 shows the differences

between the accumulated gains from the Robust Target & Manager 7 portfolio and the

Classical Target & Manager 6 portfolio (upper row). In the lower row, Figure 7 shows

the differences between the accumulated gains from Robust Minimum Risk investment &

Manager 10 and the classical version & Manager 12.

16

0 100 200 300 400 500

-2-1

01

2

Difference between the accumulated gains from the robust and classical target portfolios

0 100 200 300 400 500

-20

2

Difference between the accumulated gains from the robust and classical minimum risk portfolios

Figure 6: Difference between the accumulated gains from the best robust and best classical

portfolios.

5 Discussions

As usual, the conclusions drawn in this paper only apply to data possessing similar char-

acteristics. However, the results and discussions may shed some light on the largely

discussed topic of portfolio allocation and rebalancing strategies, and may be easily tested

and extended to other investments based on different asset characteristics (expected re-

turn, volatility and correlations). More importantly, this work has shown that the robustly

estimated pair-copulas based mean-variance inputs are more accurate thus increasing the

chances of producing a financial instrument which will truthfully yield what was expected

from it.

From the analyzes carried on it was clear the superiority of the robust portfolios. But

even a robust portfolio must be properly managed. According to the empirical analysis

of the first data set composed only by daily stock returns, for the Robust Minimum Risk

portfolio the second best manager with respect to higher accumulated gains is Manager

9 (which uses the VaR). However, Manager 9 has done 9 rebalancings, whereas Manager

12 carried on only one after one year and a half. One may be tempted to conclude that

the good performance of the winner may be just due to the changes in the economy which

17

0 100 200 300 400 500

-2-1

01

2

Difference between the accumulated gains from the robust and classical target portfolios

0 100 200 300 400 500

-0.1

5-0

.05

0.05

0.15

Difference between the accumulated gains from the robust and classical minimum risk portfolios

Figure 7: For the new data set, the difference between the accumulated gains from the best

robust and best classical portfolios.

may have taken the portfolio away from the efficient frontier. It was not the case. Figure

8 illustrates and shows the positions of the robust minimum risk portfolios from both

managers (based on weights from last rebalancing and using the entire validation period

data) along with the updated efficient frontier. We observe that the portfolios are very

close and still close to the curve, with the point risk×return from Manager 9 (triangle in

pink) even a little bit higher. The success of this procedure may be credited to the timing

of Manager 12 combined with the good stability of the robust pair-copulas method of

portfolio construction. Actually, for the second data set used based on less volatile assets,

Manager 12 was also the best option for the Classical Minimum Risk portfolio.

Manager 9 was also the second best option for the Classical Minimum Risk portfolio,

and the best one for the Robust Target. Thus, one may say that controlling the Value-at-

Risk is also an efficient strategy. Another issue is whether or not the target portfolio should

be restored to the same target. We did not address this problem, but this consideration will

certainly not change the result towards the excellent performance of the robust method.

Many other managers’ rules could be defined. Another strategy that seems promising

is a variation of the rules followed by managers 10, or 11, or 12, where the threshold

determining the need for rebalancing would not be fixed for all assets. Instead, it would

18

Risk (standard deviation)

Exp

ecte

d R

etur

n

0.020 0.022 0.024 0.026 0.028 0.030

0.00

080.

0009

0.00

100.

0011

Updated Robust Minimum Risk portfolios

(Blue: Manager 12; Pink: Manager 9)

Figure 8: Updated efficient frontier and updated robust minimum risk portfolios from

Manager 12 and Manager 9.

vary according to each asset weight or importance in the composition. In addition, it

deserves further investigation the actual value of the threshold. Other values beyond the

assumed 5% and 10% may lead to better performances.

Another important result drawn from this empirical analysis is concerned with costs.

We found that despite portfolio type, robust portfolios typically demand a smaller number

of updates lowering costs.

Finally, we found that despite the rebalancing rule, the robust portfolios always out-

perform the classical versions. Figure 9 shows the differences between the accumulated

gains from the robust and the calssical Target portfolios, having fixed the managing rule.

Given the same manager and the same portfolio target, the robust method is always su-

perior to the classical method. This is also true for the Minimum Risk portfolios, and the

outstanding performance of the Robust Minimum Risk portfolio for all fixed managers

can be seen in Figure 10.

[Figure 10 around here]

In summary, observing that the dependence between assets go beyond the linear cor-

19

0 100 200 300 400 500

-3-2

-10

12

3

M 1

0 100 200 300 400 500

-3-2

-10

12

3

M 2

0 100 200 300 400 500

-3-2

-10

12

3

M 3

0 100 200 300 400 500

-3-2

-10

12

3

M 4

0 100 200 300 400 500

-3-2

-10

12

3

M 5

0 100 200 300 400 500

-3-2

-10

12

3

M 6

0 100 200 300 400 500

-3-2

-10

12

3

M 7

0 100 200 300 400 500

-3-2

-10

12

3

M 8

0 100 200 300 400 500

-3-2

-10

12

3

M 9

0 100 200 300 400 500

-3-2

-10

12

3

M 10

0 100 200 300 400 500

-3-2

-10

12

3

M 11

0 100 200 300 400 500

-3-2

-10

12

3

M 12

Target Portfolios: difference between Robust and Classical methods and for all fixed managers

Figure 9: For the Target portfolio figure shows the differences between the robust and the

classical accumulated gains keeping fixed the manager.

Figure 10: For the Minimum Risk portfolio figure shows the differences between the robust

and the classical accumulated gains keeping fixed the manager.

relation, in this paper we proposed modeling log-returns data using robustly estimated

pair-copula models. The method is appealing simple and able to handle contaminations

that may occur when working with financial data. We illustrated the idea in the context of

emerging stock markets using two data sets composed by the most liquid Brazilian stocks,

a long-term inflation-indexed Brazilian treasury bonds index and a floating rate Brazilian

Treasury bill index. The empirical analyzes carried on in this paper indicated that for any

type of portfolio we are able to find the best manager strategy. Moreover, they indicated

that the robustly estimated pair-copulas based portfolios always outperform the classical

versions despite the managing strategy.

Methodology seems promising for any risk profile investor. The interested reader (or

investor) may easily tailor these ideas to his/her needs, repeating these exercises using

other data sets and considering other investor risk aversion levels, and even defining new

rules for managing. We are very confident that it will always be a combination of manager

and a robust portfolio outperforming its classical version. The authors will be happy to

20

compute the robust estimates and check the performance of any portfolio for any data set

the reader shall have. Future research may include the investigation of different managers’

rules for other types of portfolios.

Indeed, we have a final recommendation. If the objective is a minimum risk long

run investment the best one can do is to allocate the assets using robustly estimated

pair-copulas estimates and inspect the portfolio each 6 months, rebalancing whenever

allocations change by 10%. This will guarantee the investment characteristics and provide

the cheapest managing strategy.

Acknowledgements. The first author thanks the financial support from CNPq. Both

authors gratefully acknowledge COPPEAD research funds.

References

Aas, K., Czado, C., Frigessi, A., and Bakken, H. 2007. “Pair-copula constructions of multiple

dependence.” Insurance: Mathematics and Economics, 2, 1, 1-25.

Bedford, T., and Cooke, R. M. 2001. “Probability density decomposition for conditionally de-

pendent random variables modeled by vines.” Annals of Mathematics and Artificial Intel-

ligence, 32, 245-268.

Bedford, T., and Cooke, R. M. 2002. “Vines - a new graphical model for dependent random

variables.” Annals of Statistics, 30, 4, 10311068.

Berg, D. 2008. “Copula Goodness-of-fit testing: An overview and power comparison.” European

Journal of Finance.

Berg, D., and Aas, K. 2008. “Models for construction of multivariate dependence: A comparison

study.” European Journal of Finance.

Canela, M. A. and Collazo E. P. (2007). “Portfolio selection with skewness in emerging market

industries”, Emerging Markets Review, 8, 230250.

Chekhlov, A., Uryasev, S., and Zabarankin, M. (2003). “Portfolio Optimization with Drawdown

Constraints”. Asset and Liability Management Tools, (B. Scherer ed.), Risk Books, London,

263-278.

Demarta, S., and McNeil, A. 2005. “The t Copula and Related Copulas”. International

Statistical Review, 73(1).

Embrechts P., McNeil A., and Straumann D. 1999. “Correlation and Dependence in Risk Manage-

ment: Properties and Pitfalls.”, Preprint ETH Zurich. Available at http://www.math.ethz.ch/#embrechts;

http://citeseer.ist.psu.edu/article/embrechts99correlation.html

21

Gatzert, N., Schmeiser, H., and Schuckmann, S. 2008. “Enterprise risk management in financial

group: analysis of risk concentration and default risk.” Financial Markets and Portfolio

Management. Vol. 22, 3, 241-258.

Genest, C., and Remillard, B. 2005. “Validity of the parametric bootstrap for goodness-of-

fit testing in semiparametric models”. Technical Report G-2005-51, GERAD, Montreal,

Canada.

Genest, C., Remillard, B., and Beaudoin, D. 2007. “Omnibus goodness-of-fit tests for copulas.

A review and a power study”. Working paper, Universite Laval.

Grossman, S. J. and Zhou, Z. (1993). “Optimal investment strategies for controlling drawdowns”.

Mathematical Finance, 3, 241-276.

Hansen, B. 1994. “Autoregressive Conditional Density Estimation.” International Economic

Review, 35,3.

Jobson, J.D. and Korkie, B. (1980). ”Estimation for Markowitz Efficient Portfolios”. J. Am.

Stat.Association, 75,544-554.

Jobson, J.D. and Korkie, B. (1981). “Putting Markowitz to Work”. Journal of Portfolio

Management, 7, 4, 70-74.

Joe, H. and Xu J. (1996). “The Estimation Method of Inference Functions for Margins for

Multivariate Models.” Technical Report 166, Department of Statistics, University of British

Columbia.

Joe, H. (1997). “Multivariate Models and Dependence Concepts.” London: Chapman & Hall.

Klein, R. W. and Bawa, V. S. (1976). “The Effect of Estimation Risk on Optimal Portfolio

Choice”. Journal of Financial Economics, 3, 215-231.

Kurowicka, D., and Cooke, R. M. 2006. “Completion Problem with Partial correlation Vines”.

Linear Algebra and its Applications. Vol. 418, No. 1, pp. 188-200.

Markowitz, H.M. (1952). “Portfolio Selection”, The Journal of Finance, 7 (1): 7791.

Mendes, B. V. M., Melo, E. F. L., and Nelsen, R. B. 2007. “Robust fits for copula models”.

Communication in Statistics, 36, pp.997-1017.

Mendes, B. V. M., Semeraro, M. M. and Leal, R. P. C. (2010). “Pair-Copulas Modeling in

Finance”, Financial Markets and Portfolio Management, 24, 193-213.

Mendes, B. V. M., Melo, V. B. Accioly, 2011. “Robust Pair-Copulas Based Forecasts of Series of

Realized Volatilities”. Under Revision.

Min, A., and C. Czado 2008. “Bayesian inference for multivariate copulas using pair-copula

constructions”. Preprint, available under http://www-m4.ma.tum.de/Papers/index.html.

Nelsen, R.B. 2007. “An introduction to copulas.” Lectures Notes in Statistics. New York:

Springer.

22

Sklar, A. 1959. “Fonctions de repartition a n dimensions et leurs marges.” Publ. Inst. Statist.

Univ. Paris, 8, 229-231.

23